Deep Learning Engineering Life Cycle Problems on Edge AI Network

• This project deals with Ambient AI middleware platform that supports the edge AI devices, deep learning models and the data.

• The emergence of edge DNNs accelerating chips such as Google TPU facilitate more Deep Learning applications in the world as well as IoT environment. For example, the retail business can use "Recognition of the object and tagged bar-code", "Optimization of a show window based on user's favor" like Amazon Go, which is executed on not the central DL server but the local machines. Prof. Cha start to refer this technology making AI everywhere as an Ambient AI.

• There is a typical life cycle of the Deep Learning process (DL Life Cycle) that consists of the data preparation, model training, test, deployment and re-training as feedback. We consider the DL Life Cycle under the Ambient AI environments. There are specific cases when the edge AI devices are added to the DL Life Cycle. The figure represents 3 cases of operations between the Ambient devices and A2I middleware platform.

• Figure (a) represents a situation that a task is given under ambient AI network, which needs to find out proper model, re-train that model, and convert the model to deploy. When it comes to edge devices, there are model searching and conversion problem that builds proper models under edge AI constraints. Usually a Neural Architecture Search problems have been dealt with for more efficient computation using Reinforcement Learning approach [1,2]. We are interested in selecting model with capturing the characteristic of DNN models.

• Figure (b) represents reflecting the feedback from the edge AI devices. There are two representative problems, Federated Learning (FL) [3] and Active Learning (AL).

  • Multiple sources of training agent in FL make "statistical heterogeneity" which breaks the i.i.d. assumption, and recently it was suggested that multi-task learning approach to solve this problem [4].

  • Active Learning [5] is a special case of machine learning in which a learning algorithm can interactively query a user to label new data points with the desired outputs. Reducing the redundant teaching cost, in other words, to select a query strategy, uncertainty quantification is recently used to determine which data will be more helpful to the model [6]. Another issue is to implement the theoretical approach under the practical computing resources such as edge AI devices.

• Figure (c) represents a case where multiple edge devices operate collaboratively. We think these system as a multi-agent system (MAS), which has various open questions such as cooperation, task decomposition, consensus, network protocol [7,8].

[1] Zoph, Barret, and Quoc V. Le. "Neural architecture search with reinforcement learning." arXiv preprint arXiv:1611.01578 (2016).

[2] Zoph, Barret, et al. "Learning transferable architectures for scalable image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2018.

[3] Konečný, Jakub, et al. "Federated learning: Strategies for improving communication efficiency." arXiv preprint arXiv:1610.05492 (2016).

[4] Smith, Virginia, et al. "Federated multi-task learning." Advances in Neural Information Processing Systems. 2017.

[5] 'Active Learning (machine learning)' (18 Nov 2020‎), https://en.wikipedia.org/wiki/Active_learning_(machine_learning), 9 Dec 2020.

[6] Gal, Yarin, Riashat Islam, and Zoubin Ghahramani. "Deep bayesian active learning with image data." arXiv preprint arXiv:1703.02910 (2017).

[7] Zheng, Yuanshi, Jingying Ma, and Long Wang. "Consensus of hybrid multi-agent systems." IEEE transactions on neural networks and learning systems 29.4 (2017): 1359-1365.

[8] Kotb, Yehia, et al. "Cloud-based multi-agent cooperation for IoT devices using workflow-nets." Journal of Grid Computing 17.4 (2019): 625-650.